Beijing-based Moonshot AI apologised on Monday for rate limiting and unnecessary quota consumption after its Kimi K3.1 Pro launch, and said it will reset every subscriber’s monthly quota by 25%, which it described as one week’s usage, according to Reuters. Subscribers who have used less than 25% of their quota will have their usage counter reset to zero, the company said on its Kimi account on X. Moonshot also said it adjusted parameters to cut unnecessary subagent calls and output length. The apology comes three days after K3.1 Pro’s October 23 launch, and Reuters notes Moonshot had computing-capacity problems after the earlier K3 launch, as Chinese AI labs face computing shortages under US chip export controls.
What subscribers get back
According to Reuters, Moonshot is offering to reset the monthly quota of all subscribers by 25%. The company framed that figure as one week’s usage, so in practice subscribers get roughly a quarter of their monthly allowance returned. Anyone who has used less than a quarter of their allowance will have their usage counter reset to zero instead, wiping the small amount they have consumed so far.
The second half of the fix is technical rather than compensatory. Moonshot said it adjusted parameters to reduce unnecessary subagent calls and output length. Reuters explains the terms plainly: subagent calls are the delegation of tasks among AI agents working on a job, and output length is the total number of tokens generated in response to a prompt. Both can quietly inflate how much quota a single task consumes, so trimming them addresses the waste itself rather than only refunding it.
Reuters gave no figures for how many subscribers were affected, and no date for when the resets take effect. The reset covers all subscribers, not only those who complained, which suggests the launch problems were broad rather than confined to a few heavy users.
A launch that drew heavy demand
Kimi K3.1 Pro arrived on October 23 as a post-trained version of the Kimi K3 model released in July, carrying the same 2.8-trillion-parameter architecture with 104 billion active parameters per token and a 1-million-token context window. On Moonshot’s own numbers it scores 65.7% on the Terminal-Bench 4.0 coding test, within a point of Anthropic’s Claude Opus 5.5 at 66.4%, and 92.7% on the BenchCAD 3D-modelling test. A faster Highspeed variant runs two to three times faster, according to Moonshot.
Independent benchmarker Artificial Analysis has since ranked K3.1 Pro fourth on its Intelligence Index with 55 points, three behind Claude Opus 5.5. The model is live across the Kimi app, Kimi Work, Kimi Code and the Kimi API, with full model weights promised by November 4. That combination — near-frontier scores, a cheap API and an imminent open release — made the launch a natural magnet for heavy early traffic.
Reuters notes this is not the first time demand has outrun Moonshot’s capacity. The company previously had computing-capacity issues after K3 was launched, and the wire adds that Chinese AI labs are suffering from computing shortages following US chip export controls. Against that backdrop, a launch-day squeeze followed by a quota refund reads less as a one-off glitch and more as the recurring cost of serving frontier-scale models on constrained hardware.
Why It Matters
For Kimi subscribers, the reset returns a week’s usage, but the parameter change may matter more over time. K3.1 Pro is built for long-horizon agent work, where one request can spin up subagents that each make their own calls and generate their own output. When those background calls run longer than the job needs, quota drains without the user seeing anything extra for it. If Moonshot’s adjustment genuinely curbs that waste, subscribers should feel the difference on every agent-heavy task, not just this month’s bill.
For Moonshot, the apology is also a trust payment ahead of a bigger moment. The full K3.1 Pro weights are due by November 4, when anyone with serious hardware will be able to run a near-frontier model themselves. Kimi’s audience is professional and document-heavy rather than mass-market, and that audience is exactly the group that notices when quotas burn too fast. Handling the launch strain openly, with a refund for every subscriber, protects the goodwill Moonshot needs as it converts a popular assistant into a model company.
The wider signal is about capacity. Reuters ties Chinese labs’ computing shortages to US chip export controls, and Moonshot’s repeated launch squeezes show how that constraint bites in practice: rate limits and vanishing quotas for paying users. Until the hardware picture changes, every successful Chinese model launch carries the same risk.
Background
Moonshot AI is the Beijing startup behind the Kimi assistant, which built its name on long-context work before moving into open-weight flagship models. Kimi K3 arrived in July, K3.1 Pro followed on October 23 as its post-trained successor, and the full weights are promised by November 4. Reuters reports Moonshot also had computing-capacity problems after the K3 launch, making this week’s quota reset the second time the company has had to compensate users for demand outrunning its infrastructure.
Source: Reuters
0 Comments